1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * buffered writeback throttling. loosely based on CoDel. We can't drop |
4 | * packets for IO scheduling, so the logic is something like this: |
5 | * |
6 | * - Monitor latencies in a defined window of time. |
7 | * - If the minimum latency in the above window exceeds some target, increment |
8 | * scaling step and scale down queue depth by a factor of 2x. The monitoring |
9 | * window is then shrunk to 100 / sqrt(scaling step + 1). |
10 | * - For any window where we don't have solid data on what the latencies |
11 | * look like, retain status quo. |
12 | * - If latencies look good, decrement scaling step. |
13 | * - If we're only doing writes, allow the scaling step to go negative. This |
14 | * will temporarily boost write performance, snapping back to a stable |
15 | * scaling step of 0 if reads show up or the heavy writers finish. Unlike |
16 | * positive scaling steps where we shrink the monitoring window, a negative |
17 | * scaling step retains the default step==0 window size. |
18 | * |
19 | * Copyright (C) 2016 Jens Axboe |
20 | * |
21 | */ |
22 | #include <linux/kernel.h> |
23 | #include <linux/blk_types.h> |
24 | #include <linux/slab.h> |
25 | #include <linux/backing-dev.h> |
26 | #include <linux/swap.h> |
27 | |
28 | #include "blk-stat.h" |
29 | #include "blk-wbt.h" |
30 | #include "blk-rq-qos.h" |
31 | #include "elevator.h" |
32 | |
33 | #define CREATE_TRACE_POINTS |
34 | #include <trace/events/wbt.h> |
35 | |
36 | enum wbt_flags { |
37 | WBT_TRACKED = 1, /* write, tracked for throttling */ |
38 | WBT_READ = 2, /* read */ |
39 | WBT_KSWAPD = 4, /* write, from kswapd */ |
40 | WBT_DISCARD = 8, /* discard */ |
41 | |
42 | WBT_NR_BITS = 4, /* number of bits */ |
43 | }; |
44 | |
45 | enum { |
46 | WBT_RWQ_BG = 0, |
47 | WBT_RWQ_KSWAPD, |
48 | WBT_RWQ_DISCARD, |
49 | WBT_NUM_RWQ, |
50 | }; |
51 | |
52 | /* |
53 | * If current state is WBT_STATE_ON/OFF_DEFAULT, it can be covered to any other |
54 | * state, if current state is WBT_STATE_ON/OFF_MANUAL, it can only be covered |
55 | * to WBT_STATE_OFF/ON_MANUAL. |
56 | */ |
57 | enum { |
58 | WBT_STATE_ON_DEFAULT = 1, /* on by default */ |
59 | WBT_STATE_ON_MANUAL = 2, /* on manually by sysfs */ |
60 | WBT_STATE_OFF_DEFAULT = 3, /* off by default */ |
61 | WBT_STATE_OFF_MANUAL = 4, /* off manually by sysfs */ |
62 | }; |
63 | |
64 | struct rq_wb { |
65 | /* |
66 | * Settings that govern how we throttle |
67 | */ |
68 | unsigned int wb_background; /* background writeback */ |
69 | unsigned int wb_normal; /* normal writeback */ |
70 | |
71 | short enable_state; /* WBT_STATE_* */ |
72 | |
73 | /* |
74 | * Number of consecutive periods where we don't have enough |
75 | * information to make a firm scale up/down decision. |
76 | */ |
77 | unsigned int unknown_cnt; |
78 | |
79 | u64 win_nsec; /* default window size */ |
80 | u64 cur_win_nsec; /* current window size */ |
81 | |
82 | struct blk_stat_callback *cb; |
83 | |
84 | u64 sync_issue; |
85 | void *sync_cookie; |
86 | |
87 | unsigned int wc; |
88 | |
89 | unsigned long last_issue; /* last non-throttled issue */ |
90 | unsigned long last_comp; /* last non-throttled comp */ |
91 | unsigned long min_lat_nsec; |
92 | struct rq_qos rqos; |
93 | struct rq_wait rq_wait[WBT_NUM_RWQ]; |
94 | struct rq_depth rq_depth; |
95 | }; |
96 | |
97 | static inline struct rq_wb *RQWB(struct rq_qos *rqos) |
98 | { |
99 | return container_of(rqos, struct rq_wb, rqos); |
100 | } |
101 | |
102 | static inline void wbt_clear_state(struct request *rq) |
103 | { |
104 | rq->wbt_flags = 0; |
105 | } |
106 | |
107 | static inline enum wbt_flags wbt_flags(struct request *rq) |
108 | { |
109 | return rq->wbt_flags; |
110 | } |
111 | |
112 | static inline bool wbt_is_tracked(struct request *rq) |
113 | { |
114 | return rq->wbt_flags & WBT_TRACKED; |
115 | } |
116 | |
117 | static inline bool wbt_is_read(struct request *rq) |
118 | { |
119 | return rq->wbt_flags & WBT_READ; |
120 | } |
121 | |
122 | enum { |
123 | /* |
124 | * Default setting, we'll scale up (to 75% of QD max) or down (min 1) |
125 | * from here depending on device stats |
126 | */ |
127 | RWB_DEF_DEPTH = 16, |
128 | |
129 | /* |
130 | * 100msec window |
131 | */ |
132 | RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, |
133 | |
134 | /* |
135 | * Disregard stats, if we don't meet this minimum |
136 | */ |
137 | RWB_MIN_WRITE_SAMPLES = 3, |
138 | |
139 | /* |
140 | * If we have this number of consecutive windows with not enough |
141 | * information to scale up or down, scale up. |
142 | */ |
143 | RWB_UNKNOWN_BUMP = 5, |
144 | }; |
145 | |
146 | static inline bool rwb_enabled(struct rq_wb *rwb) |
147 | { |
148 | return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT && |
149 | rwb->enable_state != WBT_STATE_OFF_MANUAL; |
150 | } |
151 | |
152 | static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) |
153 | { |
154 | if (rwb_enabled(rwb)) { |
155 | const unsigned long cur = jiffies; |
156 | |
157 | if (cur != *var) |
158 | *var = cur; |
159 | } |
160 | } |
161 | |
162 | /* |
163 | * If a task was rate throttled in balance_dirty_pages() within the last |
164 | * second or so, use that to indicate a higher cleaning rate. |
165 | */ |
166 | static bool wb_recent_wait(struct rq_wb *rwb) |
167 | { |
168 | struct bdi_writeback *wb = &rwb->rqos.disk->bdi->wb; |
169 | |
170 | return time_before(jiffies, wb->dirty_sleep + HZ); |
171 | } |
172 | |
173 | static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, |
174 | enum wbt_flags wb_acct) |
175 | { |
176 | if (wb_acct & WBT_KSWAPD) |
177 | return &rwb->rq_wait[WBT_RWQ_KSWAPD]; |
178 | else if (wb_acct & WBT_DISCARD) |
179 | return &rwb->rq_wait[WBT_RWQ_DISCARD]; |
180 | |
181 | return &rwb->rq_wait[WBT_RWQ_BG]; |
182 | } |
183 | |
184 | static void rwb_wake_all(struct rq_wb *rwb) |
185 | { |
186 | int i; |
187 | |
188 | for (i = 0; i < WBT_NUM_RWQ; i++) { |
189 | struct rq_wait *rqw = &rwb->rq_wait[i]; |
190 | |
191 | if (wq_has_sleeper(wq_head: &rqw->wait)) |
192 | wake_up_all(&rqw->wait); |
193 | } |
194 | } |
195 | |
196 | static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw, |
197 | enum wbt_flags wb_acct) |
198 | { |
199 | int inflight, limit; |
200 | |
201 | inflight = atomic_dec_return(v: &rqw->inflight); |
202 | |
203 | /* |
204 | * For discards, our limit is always the background. For writes, if |
205 | * the device does write back caching, drop further down before we |
206 | * wake people up. |
207 | */ |
208 | if (wb_acct & WBT_DISCARD) |
209 | limit = rwb->wb_background; |
210 | else if (rwb->wc && !wb_recent_wait(rwb)) |
211 | limit = 0; |
212 | else |
213 | limit = rwb->wb_normal; |
214 | |
215 | /* |
216 | * Don't wake anyone up if we are above the normal limit. |
217 | */ |
218 | if (inflight && inflight >= limit) |
219 | return; |
220 | |
221 | if (wq_has_sleeper(wq_head: &rqw->wait)) { |
222 | int diff = limit - inflight; |
223 | |
224 | if (!inflight || diff >= rwb->wb_background / 2) |
225 | wake_up_all(&rqw->wait); |
226 | } |
227 | } |
228 | |
229 | static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct) |
230 | { |
231 | struct rq_wb *rwb = RQWB(rqos); |
232 | struct rq_wait *rqw; |
233 | |
234 | if (!(wb_acct & WBT_TRACKED)) |
235 | return; |
236 | |
237 | rqw = get_rq_wait(rwb, wb_acct); |
238 | wbt_rqw_done(rwb, rqw, wb_acct); |
239 | } |
240 | |
241 | /* |
242 | * Called on completion of a request. Note that it's also called when |
243 | * a request is merged, when the request gets freed. |
244 | */ |
245 | static void wbt_done(struct rq_qos *rqos, struct request *rq) |
246 | { |
247 | struct rq_wb *rwb = RQWB(rqos); |
248 | |
249 | if (!wbt_is_tracked(rq)) { |
250 | if (rwb->sync_cookie == rq) { |
251 | rwb->sync_issue = 0; |
252 | rwb->sync_cookie = NULL; |
253 | } |
254 | |
255 | if (wbt_is_read(rq)) |
256 | wb_timestamp(rwb, var: &rwb->last_comp); |
257 | } else { |
258 | WARN_ON_ONCE(rq == rwb->sync_cookie); |
259 | __wbt_done(rqos, wb_acct: wbt_flags(rq)); |
260 | } |
261 | wbt_clear_state(rq); |
262 | } |
263 | |
264 | static inline bool stat_sample_valid(struct blk_rq_stat *stat) |
265 | { |
266 | /* |
267 | * We need at least one read sample, and a minimum of |
268 | * RWB_MIN_WRITE_SAMPLES. We require some write samples to know |
269 | * that it's writes impacting us, and not just some sole read on |
270 | * a device that is in a lower power state. |
271 | */ |
272 | return (stat[READ].nr_samples >= 1 && |
273 | stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES); |
274 | } |
275 | |
276 | static u64 rwb_sync_issue_lat(struct rq_wb *rwb) |
277 | { |
278 | u64 now, issue = READ_ONCE(rwb->sync_issue); |
279 | |
280 | if (!issue || !rwb->sync_cookie) |
281 | return 0; |
282 | |
283 | now = ktime_to_ns(kt: ktime_get()); |
284 | return now - issue; |
285 | } |
286 | |
287 | static inline unsigned int wbt_inflight(struct rq_wb *rwb) |
288 | { |
289 | unsigned int i, ret = 0; |
290 | |
291 | for (i = 0; i < WBT_NUM_RWQ; i++) |
292 | ret += atomic_read(v: &rwb->rq_wait[i].inflight); |
293 | |
294 | return ret; |
295 | } |
296 | |
297 | enum { |
298 | LAT_OK = 1, |
299 | LAT_UNKNOWN, |
300 | LAT_UNKNOWN_WRITES, |
301 | LAT_EXCEEDED, |
302 | }; |
303 | |
304 | static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) |
305 | { |
306 | struct backing_dev_info *bdi = rwb->rqos.disk->bdi; |
307 | struct rq_depth *rqd = &rwb->rq_depth; |
308 | u64 thislat; |
309 | |
310 | /* |
311 | * If our stored sync issue exceeds the window size, or it |
312 | * exceeds our min target AND we haven't logged any entries, |
313 | * flag the latency as exceeded. wbt works off completion latencies, |
314 | * but for a flooded device, a single sync IO can take a long time |
315 | * to complete after being issued. If this time exceeds our |
316 | * monitoring window AND we didn't see any other completions in that |
317 | * window, then count that sync IO as a violation of the latency. |
318 | */ |
319 | thislat = rwb_sync_issue_lat(rwb); |
320 | if (thislat > rwb->cur_win_nsec || |
321 | (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) { |
322 | trace_wbt_lat(bdi, lat: thislat); |
323 | return LAT_EXCEEDED; |
324 | } |
325 | |
326 | /* |
327 | * No read/write mix, if stat isn't valid |
328 | */ |
329 | if (!stat_sample_valid(stat)) { |
330 | /* |
331 | * If we had writes in this stat window and the window is |
332 | * current, we're only doing writes. If a task recently |
333 | * waited or still has writes in flights, consider us doing |
334 | * just writes as well. |
335 | */ |
336 | if (stat[WRITE].nr_samples || wb_recent_wait(rwb) || |
337 | wbt_inflight(rwb)) |
338 | return LAT_UNKNOWN_WRITES; |
339 | return LAT_UNKNOWN; |
340 | } |
341 | |
342 | /* |
343 | * If the 'min' latency exceeds our target, step down. |
344 | */ |
345 | if (stat[READ].min > rwb->min_lat_nsec) { |
346 | trace_wbt_lat(bdi, lat: stat[READ].min); |
347 | trace_wbt_stat(bdi, stat); |
348 | return LAT_EXCEEDED; |
349 | } |
350 | |
351 | if (rqd->scale_step) |
352 | trace_wbt_stat(bdi, stat); |
353 | |
354 | return LAT_OK; |
355 | } |
356 | |
357 | static void rwb_trace_step(struct rq_wb *rwb, const char *msg) |
358 | { |
359 | struct backing_dev_info *bdi = rwb->rqos.disk->bdi; |
360 | struct rq_depth *rqd = &rwb->rq_depth; |
361 | |
362 | trace_wbt_step(bdi, msg, step: rqd->scale_step, window: rwb->cur_win_nsec, |
363 | bg: rwb->wb_background, normal: rwb->wb_normal, max: rqd->max_depth); |
364 | } |
365 | |
366 | static void calc_wb_limits(struct rq_wb *rwb) |
367 | { |
368 | if (rwb->min_lat_nsec == 0) { |
369 | rwb->wb_normal = rwb->wb_background = 0; |
370 | } else if (rwb->rq_depth.max_depth <= 2) { |
371 | rwb->wb_normal = rwb->rq_depth.max_depth; |
372 | rwb->wb_background = 1; |
373 | } else { |
374 | rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2; |
375 | rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4; |
376 | } |
377 | } |
378 | |
379 | static void scale_up(struct rq_wb *rwb) |
380 | { |
381 | if (!rq_depth_scale_up(rqd: &rwb->rq_depth)) |
382 | return; |
383 | calc_wb_limits(rwb); |
384 | rwb->unknown_cnt = 0; |
385 | rwb_wake_all(rwb); |
386 | rwb_trace_step(rwb, tracepoint_string("scale up" )); |
387 | } |
388 | |
389 | static void scale_down(struct rq_wb *rwb, bool hard_throttle) |
390 | { |
391 | if (!rq_depth_scale_down(rqd: &rwb->rq_depth, hard_throttle)) |
392 | return; |
393 | calc_wb_limits(rwb); |
394 | rwb->unknown_cnt = 0; |
395 | rwb_trace_step(rwb, tracepoint_string("scale down" )); |
396 | } |
397 | |
398 | static void rwb_arm_timer(struct rq_wb *rwb) |
399 | { |
400 | struct rq_depth *rqd = &rwb->rq_depth; |
401 | |
402 | if (rqd->scale_step > 0) { |
403 | /* |
404 | * We should speed this up, using some variant of a fast |
405 | * integer inverse square root calculation. Since we only do |
406 | * this for every window expiration, it's not a huge deal, |
407 | * though. |
408 | */ |
409 | rwb->cur_win_nsec = div_u64(dividend: rwb->win_nsec << 4, |
410 | divisor: int_sqrt((rqd->scale_step + 1) << 8)); |
411 | } else { |
412 | /* |
413 | * For step < 0, we don't want to increase/decrease the |
414 | * window size. |
415 | */ |
416 | rwb->cur_win_nsec = rwb->win_nsec; |
417 | } |
418 | |
419 | blk_stat_activate_nsecs(cb: rwb->cb, nsecs: rwb->cur_win_nsec); |
420 | } |
421 | |
422 | static void wb_timer_fn(struct blk_stat_callback *cb) |
423 | { |
424 | struct rq_wb *rwb = cb->data; |
425 | struct rq_depth *rqd = &rwb->rq_depth; |
426 | unsigned int inflight = wbt_inflight(rwb); |
427 | int status; |
428 | |
429 | if (!rwb->rqos.disk) |
430 | return; |
431 | |
432 | status = latency_exceeded(rwb, stat: cb->stat); |
433 | |
434 | trace_wbt_timer(bdi: rwb->rqos.disk->bdi, status, step: rqd->scale_step, inflight); |
435 | |
436 | /* |
437 | * If we exceeded the latency target, step down. If we did not, |
438 | * step one level up. If we don't know enough to say either exceeded |
439 | * or ok, then don't do anything. |
440 | */ |
441 | switch (status) { |
442 | case LAT_EXCEEDED: |
443 | scale_down(rwb, hard_throttle: true); |
444 | break; |
445 | case LAT_OK: |
446 | scale_up(rwb); |
447 | break; |
448 | case LAT_UNKNOWN_WRITES: |
449 | /* |
450 | * We started a the center step, but don't have a valid |
451 | * read/write sample, but we do have writes going on. |
452 | * Allow step to go negative, to increase write perf. |
453 | */ |
454 | scale_up(rwb); |
455 | break; |
456 | case LAT_UNKNOWN: |
457 | if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP) |
458 | break; |
459 | /* |
460 | * We get here when previously scaled reduced depth, and we |
461 | * currently don't have a valid read/write sample. For that |
462 | * case, slowly return to center state (step == 0). |
463 | */ |
464 | if (rqd->scale_step > 0) |
465 | scale_up(rwb); |
466 | else if (rqd->scale_step < 0) |
467 | scale_down(rwb, hard_throttle: false); |
468 | break; |
469 | default: |
470 | break; |
471 | } |
472 | |
473 | /* |
474 | * Re-arm timer, if we have IO in flight |
475 | */ |
476 | if (rqd->scale_step || inflight) |
477 | rwb_arm_timer(rwb); |
478 | } |
479 | |
480 | static void wbt_update_limits(struct rq_wb *rwb) |
481 | { |
482 | struct rq_depth *rqd = &rwb->rq_depth; |
483 | |
484 | rqd->scale_step = 0; |
485 | rqd->scaled_max = false; |
486 | |
487 | rq_depth_calc_max_depth(rqd); |
488 | calc_wb_limits(rwb); |
489 | |
490 | rwb_wake_all(rwb); |
491 | } |
492 | |
493 | bool wbt_disabled(struct request_queue *q) |
494 | { |
495 | struct rq_qos *rqos = wbt_rq_qos(q); |
496 | |
497 | return !rqos || !rwb_enabled(rwb: RQWB(rqos)); |
498 | } |
499 | |
500 | u64 wbt_get_min_lat(struct request_queue *q) |
501 | { |
502 | struct rq_qos *rqos = wbt_rq_qos(q); |
503 | if (!rqos) |
504 | return 0; |
505 | return RQWB(rqos)->min_lat_nsec; |
506 | } |
507 | |
508 | void wbt_set_min_lat(struct request_queue *q, u64 val) |
509 | { |
510 | struct rq_qos *rqos = wbt_rq_qos(q); |
511 | if (!rqos) |
512 | return; |
513 | |
514 | RQWB(rqos)->min_lat_nsec = val; |
515 | if (val) |
516 | RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL; |
517 | else |
518 | RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL; |
519 | |
520 | wbt_update_limits(rwb: RQWB(rqos)); |
521 | } |
522 | |
523 | |
524 | static bool close_io(struct rq_wb *rwb) |
525 | { |
526 | const unsigned long now = jiffies; |
527 | |
528 | return time_before(now, rwb->last_issue + HZ / 10) || |
529 | time_before(now, rwb->last_comp + HZ / 10); |
530 | } |
531 | |
532 | #define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) |
533 | |
534 | static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf) |
535 | { |
536 | unsigned int limit; |
537 | |
538 | if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD) |
539 | return rwb->wb_background; |
540 | |
541 | /* |
542 | * At this point we know it's a buffered write. If this is |
543 | * kswapd trying to free memory, or REQ_SYNC is set, then |
544 | * it's WB_SYNC_ALL writeback, and we'll use the max limit for |
545 | * that. If the write is marked as a background write, then use |
546 | * the idle limit, or go to normal if we haven't had competing |
547 | * IO for a bit. |
548 | */ |
549 | if ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd()) |
550 | limit = rwb->rq_depth.max_depth; |
551 | else if ((opf & REQ_BACKGROUND) || close_io(rwb)) { |
552 | /* |
553 | * If less than 100ms since we completed unrelated IO, |
554 | * limit us to half the depth for background writeback. |
555 | */ |
556 | limit = rwb->wb_background; |
557 | } else |
558 | limit = rwb->wb_normal; |
559 | |
560 | return limit; |
561 | } |
562 | |
563 | struct wbt_wait_data { |
564 | struct rq_wb *rwb; |
565 | enum wbt_flags wb_acct; |
566 | blk_opf_t opf; |
567 | }; |
568 | |
569 | static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data) |
570 | { |
571 | struct wbt_wait_data *data = private_data; |
572 | return rq_wait_inc_below(rq_wait: rqw, limit: get_limit(rwb: data->rwb, opf: data->opf)); |
573 | } |
574 | |
575 | static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data) |
576 | { |
577 | struct wbt_wait_data *data = private_data; |
578 | wbt_rqw_done(rwb: data->rwb, rqw, wb_acct: data->wb_acct); |
579 | } |
580 | |
581 | /* |
582 | * Block if we will exceed our limit, or if we are currently waiting for |
583 | * the timer to kick off queuing again. |
584 | */ |
585 | static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct, |
586 | blk_opf_t opf) |
587 | { |
588 | struct rq_wait *rqw = get_rq_wait(rwb, wb_acct); |
589 | struct wbt_wait_data data = { |
590 | .rwb = rwb, |
591 | .wb_acct = wb_acct, |
592 | .opf = opf, |
593 | }; |
594 | |
595 | rq_qos_wait(rqw, private_data: &data, acquire_inflight_cb: wbt_inflight_cb, cleanup_cb: wbt_cleanup_cb); |
596 | } |
597 | |
598 | static inline bool wbt_should_throttle(struct bio *bio) |
599 | { |
600 | switch (bio_op(bio)) { |
601 | case REQ_OP_WRITE: |
602 | /* |
603 | * Don't throttle WRITE_ODIRECT |
604 | */ |
605 | if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == |
606 | (REQ_SYNC | REQ_IDLE)) |
607 | return false; |
608 | fallthrough; |
609 | case REQ_OP_DISCARD: |
610 | return true; |
611 | default: |
612 | return false; |
613 | } |
614 | } |
615 | |
616 | static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio) |
617 | { |
618 | enum wbt_flags flags = 0; |
619 | |
620 | if (!rwb_enabled(rwb)) |
621 | return 0; |
622 | |
623 | if (bio_op(bio) == REQ_OP_READ) { |
624 | flags = WBT_READ; |
625 | } else if (wbt_should_throttle(bio)) { |
626 | if (current_is_kswapd()) |
627 | flags |= WBT_KSWAPD; |
628 | if (bio_op(bio) == REQ_OP_DISCARD) |
629 | flags |= WBT_DISCARD; |
630 | flags |= WBT_TRACKED; |
631 | } |
632 | return flags; |
633 | } |
634 | |
635 | static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio) |
636 | { |
637 | struct rq_wb *rwb = RQWB(rqos); |
638 | enum wbt_flags flags = bio_to_wbt_flags(rwb, bio); |
639 | __wbt_done(rqos, wb_acct: flags); |
640 | } |
641 | |
642 | /* |
643 | * May sleep, if we have exceeded the writeback limits. Caller can pass |
644 | * in an irq held spinlock, if it holds one when calling this function. |
645 | * If we do sleep, we'll release and re-grab it. |
646 | */ |
647 | static void wbt_wait(struct rq_qos *rqos, struct bio *bio) |
648 | { |
649 | struct rq_wb *rwb = RQWB(rqos); |
650 | enum wbt_flags flags; |
651 | |
652 | flags = bio_to_wbt_flags(rwb, bio); |
653 | if (!(flags & WBT_TRACKED)) { |
654 | if (flags & WBT_READ) |
655 | wb_timestamp(rwb, var: &rwb->last_issue); |
656 | return; |
657 | } |
658 | |
659 | __wbt_wait(rwb, wb_acct: flags, opf: bio->bi_opf); |
660 | |
661 | if (!blk_stat_is_active(cb: rwb->cb)) |
662 | rwb_arm_timer(rwb); |
663 | } |
664 | |
665 | static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio) |
666 | { |
667 | struct rq_wb *rwb = RQWB(rqos); |
668 | rq->wbt_flags |= bio_to_wbt_flags(rwb, bio); |
669 | } |
670 | |
671 | static void wbt_issue(struct rq_qos *rqos, struct request *rq) |
672 | { |
673 | struct rq_wb *rwb = RQWB(rqos); |
674 | |
675 | if (!rwb_enabled(rwb)) |
676 | return; |
677 | |
678 | /* |
679 | * Track sync issue, in case it takes a long time to complete. Allows us |
680 | * to react quicker, if a sync IO takes a long time to complete. Note |
681 | * that this is just a hint. The request can go away when it completes, |
682 | * so it's important we never dereference it. We only use the address to |
683 | * compare with, which is why we store the sync_issue time locally. |
684 | */ |
685 | if (wbt_is_read(rq) && !rwb->sync_issue) { |
686 | rwb->sync_cookie = rq; |
687 | rwb->sync_issue = rq->io_start_time_ns; |
688 | } |
689 | } |
690 | |
691 | static void wbt_requeue(struct rq_qos *rqos, struct request *rq) |
692 | { |
693 | struct rq_wb *rwb = RQWB(rqos); |
694 | if (!rwb_enabled(rwb)) |
695 | return; |
696 | if (rq == rwb->sync_cookie) { |
697 | rwb->sync_issue = 0; |
698 | rwb->sync_cookie = NULL; |
699 | } |
700 | } |
701 | |
702 | void wbt_set_write_cache(struct request_queue *q, bool write_cache_on) |
703 | { |
704 | struct rq_qos *rqos = wbt_rq_qos(q); |
705 | if (rqos) |
706 | RQWB(rqos)->wc = write_cache_on; |
707 | } |
708 | |
709 | /* |
710 | * Enable wbt if defaults are configured that way |
711 | */ |
712 | void wbt_enable_default(struct gendisk *disk) |
713 | { |
714 | struct request_queue *q = disk->queue; |
715 | struct rq_qos *rqos; |
716 | bool enable = IS_ENABLED(CONFIG_BLK_WBT_MQ); |
717 | |
718 | if (q->elevator && |
719 | test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags)) |
720 | enable = false; |
721 | |
722 | /* Throttling already enabled? */ |
723 | rqos = wbt_rq_qos(q); |
724 | if (rqos) { |
725 | if (enable && RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT) |
726 | RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT; |
727 | return; |
728 | } |
729 | |
730 | /* Queue not registered? Maybe shutting down... */ |
731 | if (!blk_queue_registered(q)) |
732 | return; |
733 | |
734 | if (queue_is_mq(q) && enable) |
735 | wbt_init(disk); |
736 | } |
737 | EXPORT_SYMBOL_GPL(wbt_enable_default); |
738 | |
739 | u64 wbt_default_latency_nsec(struct request_queue *q) |
740 | { |
741 | /* |
742 | * We default to 2msec for non-rotational storage, and 75msec |
743 | * for rotational storage. |
744 | */ |
745 | if (blk_queue_nonrot(q)) |
746 | return 2000000ULL; |
747 | else |
748 | return 75000000ULL; |
749 | } |
750 | |
751 | static int wbt_data_dir(const struct request *rq) |
752 | { |
753 | const enum req_op op = req_op(req: rq); |
754 | |
755 | if (op == REQ_OP_READ) |
756 | return READ; |
757 | else if (op_is_write(op)) |
758 | return WRITE; |
759 | |
760 | /* don't account */ |
761 | return -1; |
762 | } |
763 | |
764 | static void wbt_queue_depth_changed(struct rq_qos *rqos) |
765 | { |
766 | RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(q: rqos->disk->queue); |
767 | wbt_update_limits(rwb: RQWB(rqos)); |
768 | } |
769 | |
770 | static void wbt_exit(struct rq_qos *rqos) |
771 | { |
772 | struct rq_wb *rwb = RQWB(rqos); |
773 | |
774 | blk_stat_remove_callback(q: rqos->disk->queue, cb: rwb->cb); |
775 | blk_stat_free_callback(cb: rwb->cb); |
776 | kfree(objp: rwb); |
777 | } |
778 | |
779 | /* |
780 | * Disable wbt, if enabled by default. |
781 | */ |
782 | void wbt_disable_default(struct gendisk *disk) |
783 | { |
784 | struct rq_qos *rqos = wbt_rq_qos(q: disk->queue); |
785 | struct rq_wb *rwb; |
786 | if (!rqos) |
787 | return; |
788 | rwb = RQWB(rqos); |
789 | if (rwb->enable_state == WBT_STATE_ON_DEFAULT) { |
790 | blk_stat_deactivate(cb: rwb->cb); |
791 | rwb->enable_state = WBT_STATE_OFF_DEFAULT; |
792 | } |
793 | } |
794 | EXPORT_SYMBOL_GPL(wbt_disable_default); |
795 | |
796 | #ifdef CONFIG_BLK_DEBUG_FS |
797 | static int wbt_curr_win_nsec_show(void *data, struct seq_file *m) |
798 | { |
799 | struct rq_qos *rqos = data; |
800 | struct rq_wb *rwb = RQWB(rqos); |
801 | |
802 | seq_printf(m, fmt: "%llu\n" , rwb->cur_win_nsec); |
803 | return 0; |
804 | } |
805 | |
806 | static int wbt_enabled_show(void *data, struct seq_file *m) |
807 | { |
808 | struct rq_qos *rqos = data; |
809 | struct rq_wb *rwb = RQWB(rqos); |
810 | |
811 | seq_printf(m, fmt: "%d\n" , rwb->enable_state); |
812 | return 0; |
813 | } |
814 | |
815 | static int wbt_id_show(void *data, struct seq_file *m) |
816 | { |
817 | struct rq_qos *rqos = data; |
818 | |
819 | seq_printf(m, fmt: "%u\n" , rqos->id); |
820 | return 0; |
821 | } |
822 | |
823 | static int wbt_inflight_show(void *data, struct seq_file *m) |
824 | { |
825 | struct rq_qos *rqos = data; |
826 | struct rq_wb *rwb = RQWB(rqos); |
827 | int i; |
828 | |
829 | for (i = 0; i < WBT_NUM_RWQ; i++) |
830 | seq_printf(m, fmt: "%d: inflight %d\n" , i, |
831 | atomic_read(v: &rwb->rq_wait[i].inflight)); |
832 | return 0; |
833 | } |
834 | |
835 | static int wbt_min_lat_nsec_show(void *data, struct seq_file *m) |
836 | { |
837 | struct rq_qos *rqos = data; |
838 | struct rq_wb *rwb = RQWB(rqos); |
839 | |
840 | seq_printf(m, fmt: "%lu\n" , rwb->min_lat_nsec); |
841 | return 0; |
842 | } |
843 | |
844 | static int wbt_unknown_cnt_show(void *data, struct seq_file *m) |
845 | { |
846 | struct rq_qos *rqos = data; |
847 | struct rq_wb *rwb = RQWB(rqos); |
848 | |
849 | seq_printf(m, fmt: "%u\n" , rwb->unknown_cnt); |
850 | return 0; |
851 | } |
852 | |
853 | static int wbt_normal_show(void *data, struct seq_file *m) |
854 | { |
855 | struct rq_qos *rqos = data; |
856 | struct rq_wb *rwb = RQWB(rqos); |
857 | |
858 | seq_printf(m, fmt: "%u\n" , rwb->wb_normal); |
859 | return 0; |
860 | } |
861 | |
862 | static int wbt_background_show(void *data, struct seq_file *m) |
863 | { |
864 | struct rq_qos *rqos = data; |
865 | struct rq_wb *rwb = RQWB(rqos); |
866 | |
867 | seq_printf(m, fmt: "%u\n" , rwb->wb_background); |
868 | return 0; |
869 | } |
870 | |
871 | static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = { |
872 | {"curr_win_nsec" , 0400, wbt_curr_win_nsec_show}, |
873 | {"enabled" , 0400, wbt_enabled_show}, |
874 | {"id" , 0400, wbt_id_show}, |
875 | {"inflight" , 0400, wbt_inflight_show}, |
876 | {"min_lat_nsec" , 0400, wbt_min_lat_nsec_show}, |
877 | {"unknown_cnt" , 0400, wbt_unknown_cnt_show}, |
878 | {"wb_normal" , 0400, wbt_normal_show}, |
879 | {"wb_background" , 0400, wbt_background_show}, |
880 | {}, |
881 | }; |
882 | #endif |
883 | |
884 | static const struct rq_qos_ops wbt_rqos_ops = { |
885 | .throttle = wbt_wait, |
886 | .issue = wbt_issue, |
887 | .track = wbt_track, |
888 | .requeue = wbt_requeue, |
889 | .done = wbt_done, |
890 | .cleanup = wbt_cleanup, |
891 | .queue_depth_changed = wbt_queue_depth_changed, |
892 | .exit = wbt_exit, |
893 | #ifdef CONFIG_BLK_DEBUG_FS |
894 | .debugfs_attrs = wbt_debugfs_attrs, |
895 | #endif |
896 | }; |
897 | |
898 | int wbt_init(struct gendisk *disk) |
899 | { |
900 | struct request_queue *q = disk->queue; |
901 | struct rq_wb *rwb; |
902 | int i; |
903 | int ret; |
904 | |
905 | rwb = kzalloc(size: sizeof(*rwb), GFP_KERNEL); |
906 | if (!rwb) |
907 | return -ENOMEM; |
908 | |
909 | rwb->cb = blk_stat_alloc_callback(timer_fn: wb_timer_fn, bucket_fn: wbt_data_dir, buckets: 2, data: rwb); |
910 | if (!rwb->cb) { |
911 | kfree(objp: rwb); |
912 | return -ENOMEM; |
913 | } |
914 | |
915 | for (i = 0; i < WBT_NUM_RWQ; i++) |
916 | rq_wait_init(rq_wait: &rwb->rq_wait[i]); |
917 | |
918 | rwb->last_comp = rwb->last_issue = jiffies; |
919 | rwb->win_nsec = RWB_WINDOW_NSEC; |
920 | rwb->enable_state = WBT_STATE_ON_DEFAULT; |
921 | rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags); |
922 | rwb->rq_depth.default_depth = RWB_DEF_DEPTH; |
923 | rwb->min_lat_nsec = wbt_default_latency_nsec(q); |
924 | rwb->rq_depth.queue_depth = blk_queue_depth(q); |
925 | wbt_update_limits(rwb); |
926 | |
927 | /* |
928 | * Assign rwb and add the stats callback. |
929 | */ |
930 | mutex_lock(&q->rq_qos_mutex); |
931 | ret = rq_qos_add(rqos: &rwb->rqos, disk, id: RQ_QOS_WBT, ops: &wbt_rqos_ops); |
932 | mutex_unlock(lock: &q->rq_qos_mutex); |
933 | if (ret) |
934 | goto err_free; |
935 | |
936 | blk_stat_add_callback(q, cb: rwb->cb); |
937 | |
938 | return 0; |
939 | |
940 | err_free: |
941 | blk_stat_free_callback(cb: rwb->cb); |
942 | kfree(objp: rwb); |
943 | return ret; |
944 | |
945 | } |
946 | |